Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

CAPRI-CT: Causal Analysis and Predictive Reasoning for Image Quality Optimization in Computed Tomography

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A causal-aware ensemble of variational autoencoders can predict CT signal-to-noise ratio from images and acquisition metadata, and simulate how SNR would change under different scan settings.

desk verdict Useful predictive pipeline and public phantom dataset, but the causal inference claims don't survive contact with the paper's own equations. read the letter →

arxiv 2507.17420 v1 pith:R6MK2DEE submitted 2025-07-23 cs.CV

Sneha George Gnanakalavathy , Hairil Abdul Razak , Robert Meertens , Jonathan E. Fieldsend , Xujiong Ye , Mohammed M. Abdelsamea This is my paper
classification cs.CV
keywords CAPRI-CTcomputedtomographyimagequalitysignal-to-noiseratiocausalinferencecounterfactualvariationalautoencoderensemblelearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes CAPRI-CT, a causal-aware deep learning framework that predicts the signal-to-noise ratio (SNR) of computed tomography images from a fusion of the image itself and acquisition metadata (tube voltage, tube current, contrast agent type). The authors claim that an ensemble of variational autoencoders trained on phantom CT images can do more than regress SNR: by assuming a causal directed acyclic graph and applying do-calculus, the model estimates interventional quantities such as $P(\text{snr} \mid \text{do}(v), \text{do}(t), \text{do}(a))$, enabling what-if simulations about changing scan parameters or contrast agents without performing new physical scans. The central clinical payoff, if the causal claims hold, is that radiologists and technicians could design or adjust CT protocols by simulation rather than by repeated phantom or patient exposures. The paper reports $R^2 \approx 0.799$ on validation, and its ablation analysis identifies contrast agent type as the dominant factor influencing SNR.

What carries the argument

The load-bearing object is the causal directed acyclic graph together with the backdoor adjustment formula (Equation 10). The graph posits that tube voltage $v$, tube current $t$, and contrast agent type $a$ directly affect both the acquired image $i$ and the SNR outcome, while an unobserved latent $z$ (inferred by a VAE) also influences image and SNR. The architecture is an ensemble of VAEs: a convolutional encoder with metadata embeddings produces mean and variance for $z$, and a regression decoder maps $z$ and the explicit parameters to SNR. The causal engine is the do-calculus identity $P(\text{snr} \mid \text{do}(v=120)) = \sum_{t,a,i} P(\text{snr} \mid v=120, t, a, i) \cdot P(t, a, i)$, which the paper uses to convert the learned conditional distribution into interventional predictions without retraining.

What would settle it

Acquire a fresh set of physical phantom scans at a voltage never used in training (e.g., 120 kVp) plus the corresponding metadata, and compare the measured SNR distribution with CAPRI-CT's interventional prediction under $\text{do}(v=120)$. If the predicted mean or spread deviates beyond the ensemble's reported uncertainty, the backdoor adjustment in Equation 10 does not identify the causal effect.

Watch

Extended reading notes

Core claim

CAPRI-CT's central claim is that a single deep causal model can both predict and explain CT image quality. Given a CT image $i$ and the acquisition parameters voltage $v$, current $t$, and contrast agent $a$, an ensemble of VAEs learns a latent representation $z$ and a decoder that outputs SNR. The paper argues that the learned conditional distribution $P(\text{snr} \mid v, t, a, i)$, combined with a causal DAG in which $v$, $t$, $a$ are parents of snr (and the image is also treated as observed), identifies the interventional distribution $P(\text{snr} \mid \text{do}(v), \text{do}(t), \text{do}(a))$ via backdoor adjustment (Equation 10). Interventions are simulated by overriding one input, e.g. setting $v' = 120$ kVp and recomputing the predicted SNR; counterfactual estimates are produced by comparing the intervened prediction with the observed SNR. On a phantom dataset of 3,092 CT images, the ensemble achieves $R^2 \approx 0.799$ and the ablation analysis shows that removing contrast agent information collapses performance to $R^2 \approx 0.005$, leading the paper to conclude that contrast agent is the dominant causal factor for SNR.

Load-bearing premise

The identification of the interventional effect rests on the assumption that all parents of SNR (voltage, current, contrast agent, and the image) are observed and that the backdoor adjustment in Equation 10 is valid; but the paper's own structural equation for SNR includes an unobserved latent variable $z$ that also influences the image, so unless $z$ is not a confounder the claimed $\text{do}(v)$ estimates are conditional predictions, not causal effects.

Editorial extensions

If this is right

  • Radiologists and technicians could use CAPRI-CT to simulate how SNR would change under alternative tube voltages, currents, or contrast agents, reducing the need for repeated physical phantom scans.
  • The ablation analysis indicates that contrast agent type is the dominant causal driver of SNR in this phantom dataset, so protocol optimization should weigh contrast selection most heavily.
  • The same ensemble design could be extended to other image quality metrics such as contrast-to-noise ratio, sharpness, or artifact levels, as the paper states.
  • Because the model outputs uncertainty via ensemble variance, protocol designers can see not only a point prediction but a confidence band around a simulated SNR.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the latent $z$ captures unmeasured scanner or phantom state, the interventional simulations are likely conditional-on-$z$ predictions rather than true do-calculus estimates; this could be tested by comparing $z$ distributions across distinct scanning sessions.
  • Editorial inference: the dominance of contrast agent over voltage and current may reflect the limited range of voltage/current values in the dataset (80–140 kVp, 215/430 mAs) rather than a stable clinical ranking; re-weighting the parameter grid would clarify.
  • Editorial inference: a straightforward external validation is to hold out entire acquisition sessions and test whether the model's $\text{do}(v=120)$ prediction matches measured SNR of that session; the paper does not report such a session-level generalization test.
  • Editorial inference: the same architecture could be applied to patient CT data, where reconstruction kernels and window width/level vary, to see whether the causal graph remains identifiable beyond the phantom setting.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes CAPRI-CT, an ensemble of Variational Autoencoders that fuses CT phantom images with acquisition metadata (tube voltage, tube current, contrast agent) to predict Signal-to-Noise Ratio (SNR). The authors specify a causal DAG and deep structural equations, train an ensemble of VAEs, and report predictive performance (R² ≈ 0.799), ablations, and a series of "interventional" and "counterfactual" what-if SNR values in Table 2. The stated central contribution is that the model supports causal inference via the interventional estimand P(snr | do(v), do(t), do(a)) and enables protocol optimization without additional physical scans.

Significance. If the causal claims were valid, the ability to simulate CT protocol changes from observational data would be clinically valuable. The paper has concrete strengths: the dataset and code are publicly released, the ensemble design provides uncertainty estimates, and the predictive comparison against CNN, ResNet, and SqueezeNet is standard and reproducible. However, the load-bearing causal contribution is not supported by the manuscript's own equations. The identifiability argument contradicts the stated structural equations, the proposed intervention formula is not the truncated factorization of the stated SCM, and the counterfactual estimates lack an abduction step. The predictive component may be useful on its own, but the causal claims as written are not established.

major comments (4)
  1. [§4.2, Eqs. (8)–(10)] Equation (10) is not a valid backdoor adjustment for the SCM defined by Eqs. (1)–(3). Under that SCM, the truncated factorization for do(v = v0) involves the intervened image distribution P(i | v0, t, a) and the latent parent z of snr; it is not Σ_{t,a,i} P(snr | v0,t,a,i) P(t,a,i). Conditioning on i, which Eq. (1) makes a descendant of v, is not a backdoor adjustment, and replacing P(i | v0,t,a) with the observational P(t,a,i) does not correspond to any do-operation in this graph. Moreover, Eq. (3) makes z a parent of snr, so the statement in §4.2 that "all parents of snr (namely v, t, a, and i) are observed" is false as written. Therefore the central identifiability claim for Eq. (7) is unsupported.
  2. [§4.2, Eq. (9)] The implementation of do(v = 120) in Eq. (9) is not an intervention in the stated SCM. Equation (9) recomputes the decoder output with the voltage embedding replaced while keeping the original image i fixed, but under Eq. (1) the image is a function of v, so a genuine do(v = v0) must propagate the new voltage through the image equation (and, through Eq. (2), the latent z) before SNR is generated. Fixing i while changing v corresponds to conditioning on a descendant of the intervention, not to a do-operation. Consequently, the SN Ri values reported in Table 2 are outputs of a fitted network under modified metadata, not interventional causal estimates.
  3. [§4.2, "Counterfactual inference" and Table 2] The paper reports counterfactual SNR values (SN Rcf) in Table 2 without specifying the required abduction step. A counterfactual under an SCM requires inferring the exogenous variables u from the observed facts, then applying the intervention and computing the resulting outcome, but the manuscript describes no procedure for estimating u_i, u_z, or u_snr from the observations. Section 5 also concedes that potential confounders "could not be identified or confirmed" in this dataset. Given the unobserved latent z in Eq. (3), the counterfactual estimates in Table 2 are not justified as causal counterfactuals.
  4. [§5, Table 3 and causal perturbation analysis] The ablation study is presented as "causal structure perturbation," but removing an input variable from a predictive network only measures predictive dependence, not causal effect. The conclusion that contrast agent is the "dominant causal factor" is not established by the R² drops in Table 3, since no identification or interventional validation is provided. This is a presentation issue only if the causal claims are removed, but as written it reinforces the gap between the predictive results and the causal conclusions.
minor comments (4)
  1. [Figure 2 caption] The caption says "Causal Directed Acrylic Graph" where "Acyclic" is intended; please fix the typo.
  2. [§4.2, Eq. (7)] Equation (7) writes a joint intervention do(v), do(t), do(a), but the subsequent discussion and Table 2 typically intervene on one variable at a time. Please clarify whether the estimand of interest is the joint intervention or a set of single-variable interventions, since the adjustment formulae differ.
  3. [Table 2] Several reported SNR values are negative, and the observed SNR for the first row is -712.18. Given SNR is defined as μ/σ in Hounsfield units, a brief explanation of how negative mean intensities arise in the ROIs would help readers interpret the magnitudes and signs.
  4. [§4.1, "Causal Discovery"] The phase is called "causal discovery," but the DAG is hand-specified from assumptions rather than learned from data. Consider renaming this section to "causal graph specification" to avoid overclaiming.

Circularity Check

3 steps flagged · score 7.0 of 10

Causal claims reduce to fitted-network outputs: Eq 9's 'do(v)' re-evaluates the trained decoder with modified metadata; Eq 10 is not the truncated factorization of the stated SCM (Eq 3 gives snr latent parent z, so 'all parents observed' is false and conditioning on i opens a collider); Table 2 counterfactuals have no abduction step.

  1. fitted input called prediction [Section 4.2, Eqs 8-10 and Table 2]
    "During simulation, for instance, CAPRI-CT simulates interventions such as do(v = 120) by overriding the voltage in the model's input and recomputing the predicted SNR. ŷdo(v=120) = f (i, v′ = 120, t, a), (9) ... This embedding manipulation directly corresponds to the truncated factorization formula from Pearl's do-calculus: P (snr | do(v = 120)) = Σ_{t,a,i} P (snr | v = 120, t, a, i) · P (t, a, i), (10)."

    Eq 9 implements the do-operation as a forward pass of the fitted decoder with one metadata value overridden, leaving the image i unchanged. Under the paper's own SCM, do(v=120) must first replace i via Eq 1 (i = f_i(v,t,a,u_i)), then z via Eq 2, before snr is generated by Eq 3; Eq 9 propagates nothing. The asserted equivalence to Pearl's truncated factorization fails: Eq 10 conditions on i, a descendant of the intervened variable v, and marginalizes over the observational joint P(t,a,i), which is not the truncated factorization of Eqs 1-3. The Table 2 'intervened' and 'counterfactual' SNR values are therefore, by construction, evaluations of the fitted conditional P(snr | v,t,a,i) (Eq 8) at altered inputs: fitted outputs renamed as causal predictions, with no interventional validation.

  2. self definitional [Section 4.2 identifiability claim; Section 4.1 Eqs 2-3; Sections 5-6 limitations]
    "All parents of snr (namely v, t, a, and i) are observed, ensuring the identifiability of the causal effect via backdoor adjustment. [Section 4.1: 'The output variable signal-to-noise ratio (snr) is calculated as a function of imaging parameters ... along with the latent variable z and exogenous noise usnr. snr = fsnr(v, t, a, z, usnr) (3).']"

    The identifiability conclusion is asserted, not derived: the claimed parent set {v,t,a,i} is exactly the conditioning set of the fitted network (Eq 8), but the paper's own equations contradict it. Eq 3 makes the unobserved latent z a parent of snr, and Eq 2 defines z as latent; i is not a parent of snr in Eq 3. With z unobserved, backdoor adjustment over observed variables alone cannot identify P(snr | do(v)); moreover Eq 10 conditions on i, a descendant of v (Eq 1), which opens the collider path v → i ← z → snr. The paper's own limitations concede the gap: Section 5: 'potential confounders could not be identified or confirmed'; Section 6: the model 'does not explicitly adjust for them during training.'

1 more flagged steps
  1. other [Section 4.2 counterfactual definition; Table 2 SNR_cf column]
    "Counterfactual inference is defined as the estimation of what the outcome would have been under a different set of input conditions, given what actually happened essentially asking 'what if' questions."

    The counterfactual column SNR_cf in Table 2 is presented as causal evidence, but the paper describes no abduction step: no method for inferring the exogenous noises u_i, u_z, u_snr from the observed outcome, and no propagation of the modified inputs through Eqs 1-3. The only computational recipe given (Eq 9) is the fitted decoder evaluated with modified metadata, so nothing distinguishes the counterfactual values from in-sample predictions of the trained network. The claim that these numbers answer 'what would the outcome have been given what actually happened' is therefore unsupported by any described procedure: the counterfactuals reduce to the very fitted function that produced the factual predictions, with the causal label supplied by assumption.

full rationale

CAPRI-CT's predictive regression core is genuine, non-circular empirical work: the ensemble VAE is trained and validated against held-out SNR values, and the R² = 0.799 result with CNN/ResNet/SqueezeNet baselines stands on its own. The circularity lies in the paper's advertised novelty — causal analysis, intervention, and counterfactual what-if simulation — which reduces, by the paper's own equations, to evaluating the fitted conditional model at modified inputs. Concretely: (1) Eq 9 implements do(v=120) as f(i, v'=120, t, a), a forward pass with one metadata value overridden while the image stays fixed; nothing propagates through the stated structural equations i = f_i(v,t,a,u_i), z = f_z(i,v,t,a,u_z), snr = f_snr(v,t,a,z,u_snr), so the Table 2 'intervened' SNRs are exactly the fitted network's outputs. (2) The identifiability claim ('All parents of snr (namely v, t, a, and i) are observed') contradicts the paper's own Eq 3, whose parents of snr include the latent z; the asserted backdoor set just restates the network's conditioning set (Eq 8), and Eq 10's conditioning on i, a descendant of v, is not a valid adjustment for the stated graph. The paper itself concedes the gap in Section 5 ('potential confounders could not be identified or confirmed') and Section 6 ('it does not explicitly adjust for them during training'). (3) The SNR_cf counterfactual column is reported without any abduction step, so it too reduces to the fitted function's evaluations. No self-citations or imported uniqueness theorems are involved; the reduction is internal to the paper's own equations. Because the central causal claim reduces by construction to a fit, while the benchmarked regression retains independent empirical content, the circularity score is 7.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central causal claim rests on a hand-specified causal graph, an assumed latent variable z, and an intervention procedure that simply overrides inputs. No independent physical or formal validation is offered.

free parameters (4)
  • VAE network weights and biases = not reported
    The encoder, decoder, and embedding layers are trained on the 3092-image dataset. All counterfactual outputs in Table 2 are direct outputs of these fitted weights.
  • Metadata embedding dimensions = voltage 16, current 8, contrast agent 12
    Chosen by hand in Section 4.2. No sensitivity analysis is reported for these choices.
  • Ensemble size = 5
    Five base models are trained with different random seeds. The best model is then selected for causal inference, which is a post hoc selection.
  • Training hyperparameters = not reported
    Learning rate, batch size, latent dimension, and optimizer are not given, so the fitted function cannot be exactly reproduced.
assumptions (4)
  • ad hoc to paper The manually specified DAG (v,t,a -> i; v,t,a,z -> snr) correctly represents the true causal structure of the CT acquisition process.
    Section 4.1. The graph is asserted, not discovered or validated against physical experiments.
  • ad hoc to paper Interventions can be simulated by overriding one input value while keeping the original image i fixed.
    Section 4.2, Equation 9. This violates the structural equation i = fi(v,t,a,ui), since changing v or t should change the image; it is also not the do-calculus formula claimed.
  • domain assumption Given z and parameters, the image provides limited additional information for predicting SNR.
    Section 4.1. This assumption justifies excluding i from the SNR equation, but contradicts Equation 8, which conditions on i.
  • ad hoc to paper All parents of snr are observed, so backdoor adjustment is valid.
    Section 4.2. Equation 3 includes unobserved z as a parent of snr, making this assumption false and undermining the identifiability claim.
invented entities (1)
  • Latent variable z (shared imaging context)
    purpose: Captures unmeasured factors such as scanner characteristics and acquisition noise, and is treated as a parent of SNR in the causal model.
    z is a learned VAE latent fitted on the same data. No independent measurement, external benchmark, or falsifiable prediction outside the model is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CAPRI-CT: Causal Analysis and Predictive Reasoning for Image Quality Optimization in Computed Tomography." pith.science (2026). https://pith.science/paper/R6MK2DEE

@misc{pith2026250717420,
  author       = {Pith},
  title        = {Pith review of: CAPRI-CT: Causal Analysis and Predictive Reasoning for Image Quality Optimization in Computed Tomography},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R6MK2DEE}},
  note         = {Machine review of arXiv:2507.17420}
}
read the original abstract

In computed tomography (CT), achieving high image quality while minimizing radiation exposure remains a key clinical challenge. This paper presents CAPRI-CT, a novel causal-aware deep learning framework for Causal Analysis and Predictive Reasoning for Image Quality Optimization in CT imaging. CAPRI-CT integrates image data with acquisition metadata (such as tube voltage, tube current, and contrast agent types) to model the underlying causal relationships that influence image quality. An ensemble of Variational Autoencoders (VAEs) is employed to extract meaningful features and generate causal representations from observational data, including CT images and associated imaging parameters. These input features are fused to predict the Signal-to-Noise Ratio (SNR) and support counterfactual inference, enabling what-if simulations, such as changes in contrast agents (types and concentrations) or scan parameters. CAPRI-CT is trained and validated using an ensemble learning approach, achieving strong predictive performance. By facilitating both prediction and interpretability, CAPRI-CT provides actionable insights that could help radiologists and technicians design more efficient CT protocols without repeated physical scans. The source code and dataset are publicly available at https://github.com/SnehaGeorge22/capri-ct.

Figures

Figures reproduced from arXiv: 2507.17420 by the authors.

Figure 1
Figure 1. Sample CT image taken on perspex phantom [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. This figure represents Causal Directed Acrylic Graph (DAG) of our pro [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Overview of the CAPRI-CT architecture with an ensemble modeling ap [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Heatmap of SNR estimates - SNRobs, SNRi , and SNRcf effects, indicating the model’s ability to isolate and quantify causal impacts on SNR. Ablation Study and Causal Perturbation Analysis [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Causal Transfer in Medical Image Analysis

    cs.CV 2026-03 accept novelty 5.0 of 10

    Causal Transfer Learning unifies structural causal models, invariant risk minimisation and counterfactuals with transfer learning to produce domain-robust medical image models.

Reference graph

Works this paper leans on

17 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [1]

    L. Chen, P. Bentley, K. Mori, et al. Self-supervised learning for medical image analysis using image context restoration.Medical Image Analysis, 58:101539, 2019. 12 Gnanakalavathy, Razak, Meertens, Fieldsend, Ye & Abdelsamea

    work page 2019

  2. [2]

    Oumano, L

    M. Oumano, L. Russell, M. Salehjahromi, et al. CT imaging of gold nanoparticles in a human-sized phantom.Journal of Applied Clinical Medical Physics , 22(1):337– 342, 2021

    work page 2021

  3. [3]

    J. Li, S. Wang, K. Zhou, et al. Mambo-net: Multi-causal aware modeling backdoor- intervention optimization for medical image segmentation network.arXiv preprint, arXiv:2505.21874, 2025

    work page Pith review arXiv 2025

  4. [4]

    Castro, I

    D.C. Castro, I. Walker, and B. Glocker. Causality matters in medical imaging. Nature Communications, 11:3673, 2020

    work page 2020

  5. [5]

    N. Sani, D. Malinsky, and I. Shpitser. Explaining the behavior of black-box pre- diction algorithms with causal learning.arXiv preprint arXiv:2006.02482 , 2020

    work page Pith review arXiv 2006

  6. [6]

    Vlontzos, D

    A. Vlontzos, D. Rueckert, and B. Kainz. A review of causality for learning al- gorithms in medical image analysis.Journal of Machine Learning for Biomedical Imaging. pp 1-17 , 2022

    work page 2022

  7. [7]

    Pawlowski et al

    N. Pawlowski et al. Deep structural causal models for tractable counterfactual inference. Advances in neural information processing systems. 857-869 , 33, 2020

    work page 2020

  8. [8]

    S. Liao. Counterfactual inference on retina fundus images using deep structural causal models. Master’s thesis, 2022

    work page 2022

Show all 17 references
  1. [9]

    W. Peng, T. Xia, F.D.S. Ribeiro, et al. Latent 3D Brain MRI Counterfactual. arXiv preprint, arXiv:2409.05585, 2024

    2024

  2. [10]

    Lee, K.H

    S. Lee, K.H. Lee, J.H. Park, H.Y. Kim, Y. Choi, and K.H. Lee. Staging chest CT in patients with early-stage colon cancer: analysis of impact on survival using inverse probability weighting and causal diagram. American Journal of Roentgenology , 221(2):184–195, 2023

    2023

  3. [11]

    Caussl:Causality-inspiredsemi-supervisedlearningformedicalimage segmentation

    J.Miaoetal. Caussl:Causality-inspiredsemi-supervisedlearningformedicalimage segmentation. In Proceedings of the International Conference on Computer Vision (ICCV), 2023

    2023

  4. [12]

    Jones, D.C

    C. Jones, D.C. Castro, F.D.S. Ribeiro, O. Oktay, M. McCradden, and B. Glocker. No fair lunch: a causal perspective on dataset bias in machine learning for medical imaging. arXiv preprint arXiv:2307.16526 , 2023

    2023 arXiv

  5. [13]

    Kaddour, A

    J. Kaddour, A. Lynch, Q. Liu, et al. Causal machine learning: A survey and open problems. arXiv preprint, arXiv:2206.15475, 2022

    2022 arXiv

  6. [14]

    Dash, V.N

    S. Dash, V.N. Balasubramanian, and A. Sharma. Evaluating and mitigating bias in image classifiers: A causal perspective using counterfactuals. InProceedings of the IEEE/CVF winter conference on applications of computer vision , pages 915–924, 2022

    2022

  7. [15]

    Sahoh, K

    B. Sahoh, K. Haruehansapong, and M. Kliangkhlao. Causal artificial intelligence for high-stakes decisions: The design and development of a causal machine learning model. IEEE Access, 10:24327–24339, 2022

    2022

  8. [16]

    Huang, L

    S. Huang, L. Wang, J. Liao, and L. Liu. Multi-attentional causal intervention networksformedicalimagediagnosis. Knowledge-Based Systems,299:111993,2024

    2024

  9. [17]

    C. Li, Y. Mao, S. Liang, J. Li, Y. Wang, and Y. Guo. Deep causal learning for pancreatic cancer segmentation in CT sequences. Neural Networks, 175:106294, 2024

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.